Docs · Agent

Harness

The model is the driver, the harness is the vehicle. One agent loop owns everything; these mechanisms attach to it, each a small module, never a framework.

• Project context

  SENTINEL.md → CLAUDE.md → AGENTS.md → .sentinel/context.md are injected into the system prompt (3k chars each). Plain files, no database. The repo is the index.

• Skills

  Reusable workflows in .sentinel/skills/<name>/SKILL.md. Only names + descriptions sit in context; the full body loads when the model invokes the skill tool. Ships with reproduce-fix-verify.

• Todos

  todoWrite overwrites the FULL list every call (stops the plan drifting), persisted to .sentinel/todos.json so goals survive compaction. Statuses: pending, in_progress, completed.

• Subagents

  spawnAgent runs the same loop with fresh messages and restricted tools (depth limit 1) and returns a summary, the parent's context grows by the summary, not the transcript.

• Hooks

  PreToolUse blocks dangerous commands (rm -rf /, fork bombs) and secret writes (.env, *.pem) before permission checks. A stop hook forces a test run when files changed but no tests ran.

• Trajectories

  Every turn records JSONL to .sentinel/trajectories/<runId>.jsonl, the raw material for evals. Read failing traces to find the next behavior to gate.

Loop rules

• Consecutive read-only tool calls run concurrently (up to 10); writes run serially.
• Three edits to the same file triggers a reconsider reminder instead of another retry.
• Disable trajectory logging with SENTINEL_NO_TRAJECTORY=1.

Write your own skill

---
name: my-flow
description: When to use this workflow, in one line.
---
# My flow

1. Step one.
2. Step two.